Abstract

We report on what we believe to be the first demonstration of a fully operational optical learning machine. Learning in this machine is stochastic taking place in a self-organizing tri-layered opto-electronic neural net with plastic connectivity weights that are formed in a programmable nonvolatile spatial light modulator. The net, which can also be called a Boltzmann Learning Machine, learns by adapting its connectivity weights in accordance to environmental inputs. Learning is driven by error signals derived from statevector correlation matrices accumulated at the end of fast annealing bursts that are induced by controlled optical injection of noise into the network. Operation of the machine is made possible by two important developments in our work: Fast annealing by optically induced noisy thresholding, and stochastic learning with binary weights. Preliminary results obtained with a 24 neuron prototype show that the machine can learn, with a learning score of about 60%, to associate three 8-bit vector pairs in 10-60 minutes with relatively slow (60 msec response time) neurons and that, shifting to neurons with 1 μsec response time for example, could reduce the learning time by roughly 10⁴ times. Methods for improving the learning score are presently under investigation and initial results are encouraging. They indicate that reducing the number of vectors to be learned by the prototype from three to two and use of on-on correlations alone in computing the coincidence probabilities and corresponding error signals controlling synaptic weight modification and adjusting the number of hidden neurons can increase the learning score to 95%. A short video of the machine in operation as it learns will be shown. We close by describing methods for constructing large-scale photonic learning machines of 10³-10⁵ neurons that utilize the concepts developed and speculate on potential applications.

© 1988 Optical Society of America